This November, Mike Montemerlo's Volkswagen Passat
wagon will drive a 60-mile trek through an urban landscape located somewhere in the western U.S. But Montemerlo won't be sitting behind the wheel, nor will anyone
else. That's because Montemerlo's Passat happens to
be a special robot model, custom-developed by the Stanford University Racing Team ().
Along with robotic-vehicle researchers at other institutions worldwide, the Stanford team will compete in the DARPA (Defense
Advanced Research Projects Agency) Urban Challenge, an event sponsored by the U.S. Department of Defense. DARPA's mission is to maintain
the military's technological edge and to prevent new technology from
jeopardizing national security. To meet these goals, the agency sponsors
revolutionary, high-payoff research ventures such as the Urban Challenge
that can lead to innovative new military capabilities as well as leading-edge consumer or business products and services.
The Urban Challenge's roots lay in a 2001 Congressional mandate that requires one-third of the
nation's combat ground vehicles to be unmanned by
2015. To help reach this goal, Congress authorized the Department of Defense and DARPA to award funding and cash prizes to Urban Challenge participants. Eleven teams received seed money of up to $1 million for development while another 78 are funded either by themselves or by corporate sponsors. The competition's top three finishers can look forward to receiving cash prizes of $2 million, $1 million, and $500,000,
respectively.
Self Control
The Urban Challenge is designed to test the safety of autonomous vehicles as they interact with other vehicles and the local environment. Come
this November, dozens of vehicles will be placed onto a course that simulates real-world city streets.
"All of the participants' vehicles will be on the course at the same time,
with some additional traffic vehicles that we provide," says Norman
Whitaker, the DARPA's Urban Challenge program manager. "It really tests
lots of different dimensions of the technology... both their ability to sense
and react correctly in traffic as well as their ability to be able to adapt to
changing [road] conditions."
To succeed in the Urban Challenge, competing teams' entries must perform like vehicles controlled by human drivers and safely conduct simulated battlefield supply missions on the 60-mile course. The vehicles must
obey strict rules while merging into traffic, navigating traffic circles, negotiating busy intersections, and avoiding obstacles before finishing under six
hours. The competition's location will be announced in August.
Building A Robot Car
The strongest Urban Challenge competitors are teamed with major
automakers and other sponsors. Volkswagen, in Stanford's case, provides research expertise and funding. "Volkswagen does all the retrofitting of the vehicle and all the vehicle electronics," says Montemerlo,
the team's software lead and a senior research engineer in the Stanford
Artificial Intelligence Lab. "At Stanford, we concentrate on what we do
best, which is the software and sensing."
Stanford's robo-Passat, nicknamed "Junior," is like a NASCAR racer in
that the vehicle is so highly modified it bears little more than a passing
exterior resemblance to its street counterpart. Junior's steering, throttle,
and brakes were all modified by engineers at the Volkswagen of America
Electronics Research Laboratory in Palo Alto, Calif., to be completely computer-controllable (). The engineers also created custom mountings
for a bevy of sophisticated sensors.
Junior's standard equipment includes
a range-finding laser array that spins to
provide a 360° 3D view of the surrounding environment, nearly in real-time. Six
video cameras that accompany the
array "see" all around the car. Junior
also uses bumper-mounted lasers,
radar, GPS receivers, and inertial navigation hardware to collect data about
where it is and what's nearby.
According to Montemerlo, developing
and installing the control and monitoring
hardware was the easy part. Teaching
Junior how to safely drive itself is proving
to be far more tricky, requiring programmers not only to instruct the robot on
how to control itself, but also how to
anticipate changes in traffic conditions.
"You have to make models of other vehicles and predict how you think they're going to behave," Montemerlo says.
Junior's custom-coded software modules include a planner (for making
decisions and choosing routes), a mapper (which transforms sensor readings into environment understanding), a localizer (which refines the GPS
position by visual observations), and a controller (which actuates planner
decisions). The entire system runs on rack-mounted servers equipped with
Intel Core 2 Duo processors. Data is processed from the vehicle's instruments as frequently as 200 times per second, Montemerlo notes.
But even with all of its cutting-edge hardware and software support,
Junior remains at a cognitive disadvantage compared to even a novice
human driver. "The robot actually has a lot less information about what other vehicles might be doing," Montemerlo says. "A robot certainly won't be
able to interpret other cars' turn signals, for instance." Nonetheless, he's
hopeful that Junior will perform well this November. "We do our best to train
for the worst possibility and hope that it works out," he says.